Selected work

Representative Papers

ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios

Jan 13, 2026

Existing retrieval-augmented generation (RAG) evaluation benchmarks struggle to address real-world challenges such as multi-document synthesis, visual understanding, and fine-grained source attribution. To bridge this gap, this work introduces the first comprehensive multimodal RAG benchmark that integrates visual content, cross-document reasoning, and multilingual support. The benchmark comprises 26,000 visually rich document pages spanning ten specialized domains and 3,099 human-verified queries, accompanied by high-quality annotations for retrieval relevance, bounding box localization, and reference answers. Systematic evaluation reveals that vision-aware retrievers significantly outperform purely text-based approaches, and late interaction with re-ranking further enhances performance. Nevertheless, current models still exhibit notable deficiencies in interpreting non-textual elements, answering open-ended questions, and achieving fine-grained visual grounding.

3 citations1 influentialRead paper

Working Notes on Late Interaction Dynamics: Analyzing Targeted Behaviors of Late Interaction Models

Mar 27, 2026

This study systematically investigates length bias in late-interaction retrieval models and the characteristics of similarity distributions beyond the MaxSim pooling mechanism. Through fine-grained analysis of prominent multi-vector late-interaction models on the NanoBEIR benchmark, the work provides the first empirical evidence that both causal and bidirectional architectures are susceptible to length bias. Furthermore, it reveals that suboptimal matching signals—those beyond the top-1 similar token—contribute negligible retrieval gains, thereby validating the sufficiency of MaxSim in capturing effective semantic matches. These findings not only confirm the practical existence of length bias but also offer theoretical support, from the perspective of similarity distribution, for the widespread effectiveness of the MaxSim operator in dense retrieval systems.

0 citationsRead paper

Experiential Reflective Learning for Self-Improving LLM Agents

Mar 25, 2026

This work addresses the limited environmental adaptability and inability of large language model (LLM) agents to reuse past experiences, which often forces them to solve tasks from scratch repeatedly. To overcome this, the authors propose the Experience Reflection Learning (ERL) framework, which, after a single task attempt, retrospectively analyzes the execution trajectory to automatically distill transferable heuristic rules. During subsequent tasks, ERL selectively retrieves relevant rules and injects them into the reasoning context to guide decision-making. This approach represents the first method to achieve cross-task knowledge abstraction through reflection and enables efficient self-improvement. Evaluated on the Gaia2 benchmark, ERL improves task success rates by 7.8% over the ReAct baseline and significantly outperforms existing experience-based learning methods, demonstrating the effectiveness of heuristic abstraction and selective retrieval.

0 citationsRead paper

ModernVBERT: Towards Smaller Visual Document Retrievers

Oct 01, 2025

Existing vision-language models (VLMs) for visual document retrieval suffer from reliance on large-scale VLM fine-tuning, resulting in limited performance and poor computational efficiency. To address this, we propose ModernVBERT—a compact 250M-parameter multimodal retrieval framework. Methodologically, it innovatively integrates high-resolution image inputs, cross-modal attention masking, modality-aligned data augmentation, and a late-interaction contrastive learning objective, departing from conventional end-to-end fine-tuning paradigms. Our key contribution is achieving superior retrieval performance with significantly reduced model size: on standard document retrieval benchmarks, ModernVBERT substantially outperforms state-of-the-art models with ten times more parameters, simultaneously improving accuracy and inference efficiency. The code and pretrained models are publicly available.

0 citationsRead paper

Should We Still Pretrain Encoders with Masked Language Modeling?

Jul 01, 2025

This work investigates the relative merits of masked language modeling (MLM) versus causal language modeling (CLM) for text representation pretraining in the large-model era. Method: We conduct a large-scale, controlled empirical study across 30 models spanning 210M to 1B parameters, involving over 15,000 fine-tuning evaluations. We further propose a novel two-stage “CLM→MLM” training strategy and assess low-cost adaptation of existing CLM models to MLM objectives. Contribution/Results: We find that MLM consistently achieves superior downstream performance, while CLM exhibits higher data efficiency and fine-tuning stability. The two-stage strategy attains optimal performance under fixed compute budgets, and CLM-to-MLM adaptation significantly reduces training cost. Crucially, MLM’s advantage stems not from confounding factors such as scale or dataset bias, but from the intrinsic suitability of its objective for representation learning; CLM’s value lies in training efficiency and robustness—indicating complementarity, not mutual exclusivity, between the two paradigms.

0 citationsRead paper

Recent publications

Latest Papers

Working Notes on Late Interaction Dynamics: Analyzing Targeted Behaviors of Late Interaction Models

Mar 27, 2026

This study systematically investigates length bias in late-interaction retrieval models and the characteristics of similarity distributions beyond the MaxSim pooling mechanism. Through fine-grained analysis of prominent multi-vector late-interaction models on the NanoBEIR benchmark, the work provides the first empirical evidence that both causal and bidirectional architectures are susceptible to length bias. Furthermore, it reveals that suboptimal matching signals—those beyond the top-1 similar token—contribute negligible retrieval gains, thereby validating the sufficiency of MaxSim in capturing effective semantic matches. These findings not only confirm the practical existence of length bias but also offer theoretical support, from the perspective of similarity distribution, for the widespread effectiveness of the MaxSim operator in dense retrieval systems.

0 citationsRead paper

Experiential Reflective Learning for Self-Improving LLM Agents

Mar 25, 2026

This work addresses the limited environmental adaptability and inability of large language model (LLM) agents to reuse past experiences, which often forces them to solve tasks from scratch repeatedly. To overcome this, the authors propose the Experience Reflection Learning (ERL) framework, which, after a single task attempt, retrospectively analyzes the execution trajectory to automatically distill transferable heuristic rules. During subsequent tasks, ERL selectively retrieves relevant rules and injects them into the reasoning context to guide decision-making. This approach represents the first method to achieve cross-task knowledge abstraction through reflection and enables efficient self-improvement. Evaluated on the Gaia2 benchmark, ERL improves task success rates by 7.8% over the ReAct baseline and significantly outperforms existing experience-based learning methods, demonstrating the effectiveness of heuristic abstraction and selective retrieval.

0 citationsRead paper

ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios

Jan 13, 2026

Existing retrieval-augmented generation (RAG) evaluation benchmarks struggle to address real-world challenges such as multi-document synthesis, visual understanding, and fine-grained source attribution. To bridge this gap, this work introduces the first comprehensive multimodal RAG benchmark that integrates visual content, cross-document reasoning, and multilingual support. The benchmark comprises 26,000 visually rich document pages spanning ten specialized domains and 3,099 human-verified queries, accompanied by high-quality annotations for retrieval relevance, bounding box localization, and reference answers. Systematic evaluation reveals that vision-aware retrievers significantly outperform purely text-based approaches, and late interaction with re-ranking further enhances performance. Nevertheless, current models still exhibit notable deficiencies in interpreting non-textual elements, answering open-ended questions, and achieving fine-grained visual grounding.

3 citations1 influentialRead paper

ModernVBERT: Towards Smaller Visual Document Retrievers

Oct 01, 2025

Existing vision-language models (VLMs) for visual document retrieval suffer from reliance on large-scale VLM fine-tuning, resulting in limited performance and poor computational efficiency. To address this, we propose ModernVBERT—a compact 250M-parameter multimodal retrieval framework. Methodologically, it innovatively integrates high-resolution image inputs, cross-modal attention masking, modality-aligned data augmentation, and a late-interaction contrastive learning objective, departing from conventional end-to-end fine-tuning paradigms. Our key contribution is achieving superior retrieval performance with significantly reduced model size: on standard document retrieval benchmarks, ModernVBERT substantially outperforms state-of-the-art models with ten times more parameters, simultaneously improving accuracy and inference efficiency. The code and pretrained models are publicly available.

0 citationsRead paper

Should We Still Pretrain Encoders with Masked Language Modeling?

Jul 01, 2025

This work investigates the relative merits of masked language modeling (MLM) versus causal language modeling (CLM) for text representation pretraining in the large-model era. Method: We conduct a large-scale, controlled empirical study across 30 models spanning 210M to 1B parameters, involving over 15,000 fine-tuning evaluations. We further propose a novel two-stage “CLM→MLM” training strategy and assess low-cost adaptation of existing CLM models to MLM objectives. Contribution/Results: We find that MLM consistently achieves superior downstream performance, while CLM exhibits higher data efficiency and fine-tuning stability. The two-stage strategy attains optimal performance under fixed compute budgets, and CLM-to-MLM adaptation significantly reduces training cost. Crucially, MLM’s advantage stems not from confounding factors such as scale or dataset bias, but from the intrinsic suitability of its objective for representation learning; CLM’s value lies in training efficiency and robustness—indicating complementarity, not mutual exclusivity, between the two paradigms.

0 citationsRead paper